When Sales changes a price in one file and Operations updates delivery times in another, who signs off before an AI agent uses the new information with customers? It should not be whoever happened to upload the document. The business owner of the fact should approve the change, with a clear effective date and version record.
This is different from collecting information for an initial launch. Our business data preparation guide covers what to gather. This article addresses who can approve a change and when customers should hear it.
Updating a file is not the same as authorizing a decision
Consider an Egyptian service company running a seasonal offer. Marketing has a document saying the promotion lasts until month-end. Sales has already told customers it has ended. Finance has not approved the next discount. An agent that can retrieve all three documents may produce an answer supported by a file but not by the current business decision.
Separate the process into three steps: propose the change, approve it, and test the resulting customer answer. This does not require an AI platform with a built-in approval engine. A controlled spreadsheet and named owners can be enough to start.
Keep a source register for decisions that matter
For each consequential fact, capture the source system or file, business owner, authorized approver, version and effective date, review status, and the products or branches affected.
A simple status flow is proposed, under review, approved, or retired. You do not need this overhead for every line of marketing copy. Start with prices, discounts, operating hours, availability, refunds, and contractual commitments. These are the details where a wrong answer can create an operational obligation.
Example: delivery times change at one branch
Operations proposes changing one branch's standard delivery day from Tuesday to Thursday. Sales checks orders where Tuesday was already promised. Support prepares an explanation for affected customers. The authorized manager approves the new schedule and records when it takes effect.
A question about an existing order is not necessarily answered by the new general branch schedule. Distinguish current public information from a specific prior commitment that needs a human review.
Test the change before it reaches a live conversation
As an initial internal exercise, use at least five representative conversations. Five is a starting suggestion, not a validated universal threshold. Ask about the new schedule, the old one, a customer with an earlier commitment, an incomplete question, and a request covering two branches.
Record whether the answer cites the approved fact, admits uncertainty where appropriate, and routes exceptions to a person. When it fails, determine whether the problem is source quality, business instructions, or retrieval. Rewriting a prompt cannot make conflicting business records authoritative.
The NIST AI Risk Management Framework treats governance, measurement and ongoing management as complementary activities. It is voluntary general guidance, not a claim of certification or an Egypt-specific legal requirement.
When should the agent stop answering automatically?
If two supposedly approved sources conflict, a new price is still awaiting authorization, or a special promise is missing from the record, the safer response is to say that the detail needs checking. Specify which team owns the exception, what context should be handed over, and how the customer will receive the confirmed answer.
A useful first-week exercise
Choose three high-risk knowledge categories and assign an owner and approver to each. Review the ten most recent conversations involving prices, dates or policies. Turn each conflict into a documented correction and a new test. Expand only where recurring errors justify the effort.
With Mr. AI, teams can supply company knowledge, configure rules, test replies and review or take over conversations. Internal approval authority remains your business process. We are not claiming an unpublished built-in approval workflow. To review knowledge setup and human handoff for your team, book a demo.
